Computer School, Wuhan University, Wuhan 430072, China.
Department of Computer Science, Binghamton University, Binghamton 13902, USA.
Neural Netw. 2018 Oct;106:152-167. doi: 10.1016/j.neunet.2018.07.007. Epub 2018 Jul 18.
As one of the most common user interactive behaviors in many social media services, mention plays a significant role in both user interaction and information cascading. While an increasing line of work has focused on analyzing the mention mechanism for information diffusion, the essential problem of mentionee recommendation from the perspective of common users, i.e., how to find mentionees (mentioned users) who are most likely to be notified by a mentioner (mentioning user) for knowing a post, has been seldom investigated. This paper aims to develop personalized recommendation techniques to automatically generate mentionees when a user intends to mention others in a post. After analyzing real-world social media datasets we observe that users' mention behaviors are influenced by not only the semantic but also the spatial context factors of their mentioning activities, which motivate the needs for spatial context-aware user mention behavior modeling. In light of these, we proposed a joint probabilistic model, named Spatial COntext-aware Mention behavior Model (SCOMM), to simulate the process of generating users' location-tagged mentioning activities. By exploiting the semantic and spatial context factors in a unified way, SCOMM was able to reveal users' preferences behind their mention behaviors and provide a knowledge model for accurate mentionee recommendations. Furthermore, we designed an Item-Attribute Pruning (IAP) algorithm to overcome the curse of dimensionality and facilitate online top-k query performance. Extensive experiments were conducted on two real-world datasets to evaluate the performance of our methods. The experimental results demonstrated the superiority of our approach by making more effective and efficient recommendations compared with other state-of-the-art methods.
作为许多社交媒体服务中最常见的用户交互行为之一，提及在用户交互和信息级联中都起着重要作用。虽然越来越多的工作集中在分析提及机制以进行信息扩散，但从普通用户的角度来看，提及对象（被提及用户）推荐的基本问题，即如何找到最有可能被提及者（提及用户）通知以了解帖子的提及对象，尚未得到充分研究。本文旨在开发个性化推荐技术，以便在用户打算在帖子中提及其他人时自动生成提及对象。在分析真实的社交媒体数据集后，我们观察到用户的提及行为不仅受到语义因素的影响，还受到提及活动的空间上下文因素的影响，这促使我们需要进行空间上下文感知的用户提及行为建模。有鉴于此，我们提出了一个联合概率模型，称为空间上下文感知提及行为模型（SCOMM），以模拟用户生成带有位置标签的提及活动的过程。通过以统一的方式利用语义和空间上下文因素，SCOMM 能够揭示用户提及行为背后的偏好，并为准确的提及对象推荐提供知识模型。此外，我们设计了一个项目-属性剪枝（IAP）算法来克服维度灾难并提高在线 top-k 查询性能。我们在两个真实数据集上进行了广泛的实验，以评估我们方法的性能。实验结果表明，与其他最先进的方法相比，我们的方法通过更有效地推荐，具有优越性。
